Research On The Broad Incremental Learning System Based On Mutual Information Neural Network

With the rapid development of information technology and cloud computing technology,the mode of centralized processing user data in the cloud is becoming more and more popular.Online data flow has the characteristics of mass,time-varying and rapidity.Machine learning needs a lot of computing and storage resources,which can’t meet the needs of online scene.However,because incremental learning method has the characteristics of fast learning and no need to retrain the model,it is one of the effective methods to solve this problem.But the existing algorithms have some disadvantages,such as weak representation ability and inability to resist data distribution shift.Therefore,this paper proposes a broad learning incremental model based on mutual information neural networks.First,we propose a new unsupervised feature engineering construction algorithm,which called mutual information neural network(MNN).It is the first deep network model that uses mutual information function as loss function alone,and contains two discriminator networks to simplify the calculation of mutual information between highdimensional vectors.We compare the performance of this method with existing algorithms on 4 classifiers.Experimental results show that,compared with existing algorithms,the construction time of MNN is reduced by 60-80%,and the average accuracy is improved by 1.5-4.5%.So we can draw a conclusion that MNN can quickly construct feature engineering,which is suitable for online incremental scene.Secondly,in order to solve the problem that the existing methods can’t resist the data distribution shift.This paper proposes a broad incremental system based on mutual information neural network,which uses MNN and its linear mapping set to replace the original network structure.At the same time,a random attenuation factor assignment scheme is proposed to improve strong linear correlation between weights.Then we simulate two online incremental scenes,human activity recognition and image transmission classification,and select 3 incremental algorithms for performance comparison.The results show that,the accuracy gap between this system and the existing algorithms is increased from 0.8-1%to 4-5%.Its time consumption is about 30%higher than broad learning,which is 40-80%of other algorithms.It can be concluded that the improved model makes use of the bearable time consumption to boost the classification and anti-overfitting ability.Our research shows that the proposed incremental system is superior to existing models in terms of model representation ability,resistance to data distribution shift and model adjustment speed.This work has the significant contribution for online incremental learning scene which needs to process the vast amounts of complex data.